Computers & Geosciences: "GraphRAG and LLM-driven semantic exploration of critical mineral data"

Authors: 

• Armita Davarpanah - Environmental and Health Sciences Department, Spelman College, Atlanta, GA, 30314, USA
• Hassan A. Babaie - Department of Geosciences, Georgia State University, Atlanta, GA, 30302, USA
• W. Crawford Elliott - Department of Geosciences, Georgia State University, Atlanta, GA, 30302, USA
• Yuanzhi Tang - School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA
• Paul A. Schroeder - Department of Geology, Franklin College of Arts and Sciences, University of Georgia, Athens, GA, 30602, USA

Abstract:

The identification of patterns in the distribution and relationships of critical minerals is fundamentally challenging due to the heterogeneity, scale, and semantic complexity of geoscience data. This study introduces a deterministic-first hybrid AI architecture that extends beyond standard retrieval-augmented generation (RAG) by explicitly separating symbolic reasoning, neural semantic retrieval, and generative synthesis into complementary and interpretable reasoning pathways. Unlike conventional RAG systems that rely primarily on text-based retrieval and opaque large language model (LLM) inference, our framework integrates ontology-grounded knowledge graph querying for authoritative reasoning, vector-based semantic retrieval for contextual generalization, and LLM synthesis for natural-language interaction. The architecture is instantiated through the construction of a large-scale Critical Minerals Knowledge Graph (CMKG) derived from the Critical Minerals in Ores (CMiO) dataset, comprising more than 29,000 sample records and implemented in Neo4j using a strict ontology and modular batching pipeline. Deterministic Cypher queries provide precise answers when explicit semantic relationships are available, while dense vector embeddings indexed with FAISS supply relevant context when exact matches are unavailable. An interactive Jupyter interface exposes these reasoning modes side-by-side to support transparency and interpretability. Evaluation using structured benchmark queries demonstrates that deterministic graph querying provides reliable and explainable results for structured scientific questions, while semantic and generative components contribute primarily to contextual support and interpretation. By unifying symbolic and neural reasoning within a controlled GraphRAG framework, this work contributes a generalizable AI architecture for transparent question answering over complex scientific knowledge graphs, with direct implications for critical mineral analysis and beyond.

Highlights

• Developed an ontology-driven knowledge graph for large-scale critical mineral datasets.
• Integrated deterministic graph querying with semantic vector retrieval for geoscience analysis.
• Enabled interpretable semantic exploration of heterogeneous mineral sample and deposit data.
• Supported transparent querying across sample-, deposit-, and region-level geological scales.
• Demonstrated scalable semantic data integration using Neo4j and vector-based indexing.